Tools
Memory Spark
By OpenClaw memory plugin: LanceDB + NVIDIA Spark GPU embeddings for autonomous AI agents
Install
npm install @lancedb/lancedb`Configuration Example
{
"plugins": {
"slots": { "memory": "memory-spark" },
"allow": ["memory-spark"],
"entries": { "memory-spark": { "enabled": true } },
"load": { "paths": ["~/.openclaw/extensions/memory-spark"] }
}
}README
# memory-spark
An OpenClaw memory plugin with LanceDB vector storage and NVIDIA Spark GPU-accelerated embeddings. Drop-in replacement for `memory-core` with autonomous recall, capture, file ingestion, and hybrid search.
## Features
- **Auto-recall** — injects relevant memories before every agent turn (`before_prompt_build` hook)
- **Auto-capture** — extracts and stores facts/preferences from conversations (`agent_end` hook)
- **File watcher** — auto-indexes your workspace memory dirs, sessions, and watched paths on boot and on change
- **PDF / DOCX / audio ingest** — parse and embed documents via `pdftotext`, `mammoth`, and STT
- **Hybrid search + rerank** — vector search + BM25 full-text, reranked by Cohere-compatible cross-encoder
- **NVIDIA Spark GPU embeddings** — `nvidia/llama-embed-nemotron-8b` (4096d), 0.25s/call on Blackwell GPU
- **Serialized embed queue** — retry with exponential backoff, health monitoring, auto-cooldown
- **Dimension lock** — refuses to start if provider/dims mismatch, preventing silent vector corruption
- **Zero-touch** — configure once, everything else is automatic
## Architecture
```
memory-spark/
├── index.ts # Plugin entry: register(), tool handlers, hooks
├── src/
│ ├── config.ts # Config schema + defaults + .env loader
│ ├── manager.ts # MemorySearchManager: search, store, forget
│ ├── auto/
│ │ ├── recall.ts # before_prompt_build hook — injects memories
│ │ └── capture.ts # agent_end hook — classifies + stores new facts
│ ├── embed/
│ │ ├── provider.ts # EmbedProvider: Spark → OpenAI → Gemini fallback
│ │ ├── queue.ts # EmbedQueue: serial requests, retry, health tracking
│ │ ├── chunker.ts # Markdown-aware token-bounded chunking with overlap
│ │ └── dims-lock.ts # Dimension lock: validates provider/model/dims on startup
│ ├── ingest/
│ │ ├── pipeline.ts # extract → chunk → NER → embed → upsert
│ │ ├── watcher.ts # Chokidar watcher + boot pass across all agent workspaces
│ │ ├── workspace.ts # Agent workspace discovery, session scanning
│ │ └── parsers.ts # PDF (pdftotext+OCR), DOCX (mammoth), audio (STT), text
│ ├── classify/
│ │ ├── ner.ts # NER via Spark 18112 (dslim/bert-base-NER)
│ │ └── zero-shot.ts # Zero-shot classifier via Spark 18113 (bart-large-mnli)
│ ├── rerank/
│ │ └── reranker.ts # Cohere-compatible reranker via Spark 18096
│ └── storage/
│ ├── backend.ts # StorageBackend interface
│ └── lancedb.ts # LanceDB: connect, upsert (mergeInsert), vectorSearch, FTS, delete
```
## Requirements
- [OpenClaw](https://openclaw.ai) — the plugin host
- [LanceDB](https://lancedb.github.io/lancedb/) — `npm install @lancedb/lancedb`
- `pdftotext` (poppler-utils) — for PDF parsing
- NVIDIA Spark microservices (optional but recommended):
- **Embed** — port 18091 (`nvidia/llama-embed-nemotron-8b`, 4096d)
- **Reranker** — port 18096 (`nvidia/llama-nemotron-rerank-1b-v2`)
- **NER** — port 18112 (`dslim/bert-base-NER`)
- **Zero-shot** — port 18113 (`facebook/bart-large-mnli`)
- **OCR** — port 18097
- **STT** — port 18094
Falls back to OpenAI or Gemini embeddings if Spark is unavailable.
## Setup
### 1. Install
```bash
cd ~/.openclaw/extensions
git clone https://github.com/kleinpanic/memory-spark
cd memory-spark && npm install && npm run build
```
### 2. Configure OpenClaw
In `~/.openclaw/openclaw.json`:
```json
{
"plugins": {
"slots": { "memory": "memory-spark" },
"allow": ["memory-spark"],
"entries": { "memory-spark": { "enabled": true } },
"load": { "paths": ["~/.openclaw/extensions/memory-spark"] }
}
}
```
### 3. Environment
Set your Spark bearer token in `~/.openclaw/.env`:
```
SPARK_BEARER_TOKEN=your_token_here
```
### 4. Config (openclaw.json)
```json
{
"memory": {
"backend": "lancedb",
"lancedb": { "uri": "~/.openclaw/data/memory-spark/lancedb" },
"embed": {
"provider": "spark",
"model": "nvidia/llama-embed-nemotron-8b",
"baseUrl": "http://localhost:18091/v1"
},
"recall": {
"enabled": true,
"topK": 5,
"minScore": 0.3,
"maxTokens": 1500
},
"capture": {
"enabled": true,
"categories": ["fact", "preference", "decision"]
},
"watch": {
"enabled": true,
"indexOnBoot": true,
"debounceMs": 2000,
"paths": []
}
}
}
```
## Tools
Exposes four tools to agents:
| Tool | Description |
|------|-------------|
| `memory_search` | Hybrid vector + FTS search with reranking |
| `memory_get` | Retrieve a specific chunk by ID |
| `memory_store` | Store a fact/note manually |
| `memory_forget` | Delete memories by path or query |
## License
MIT
## Acknowledgments
Built on [LanceDB](https://lancedb.github.io/lancedb/), [NVIDIA Nemotron](https://developer.nvidia.com/nemotron), and [OpenClaw](https://openclaw.ai).
tools
Comments
Sign in to leave a comment